So - how many IEEE conferences are there!? Also things change, it was much easier to get into Neurips 7 years ago... but it's not necessarily the case that the papers from this year will have as much impact or be as good as the papers from then. And as Neurips itself showed, the peer review process is somewhat random - papers are rejected by different panels meaning that getting in is likely an achievement but also possibly a bit lucky. I don't think that evaluating the papers based on where they were published is a good way to allocate public money.
As a strong concrete example, paper 13 from Ferguson's group at Imperial is probably one of the most important documents for the last 20 years (if you live in the UK or France), it was "published" on a website...
The REF is evaluated by humans. I take 2/3 weeks to read a paper. For good papers I might work on them for 6 weeks + to really get into the technique. How can a REF reviewer consume 5 papers to evaluate an academic? How can they consume 200 papers from 40 academics?
The right thing would be to have the department sumbit it's top 5 papers - 3 reviewers could really see what is going on is a department then.